And lol yes! if you knew the dude who wanted to try that, you would know what an “extrovert” was. He was the most extroverted engineer dude I have met to date.

I even worked on a team where one engineer loved the idea and championed doing it as early as 2013. Most of these IDE extensions didn't exist back then. He ended up picking a buddy, and they moved to the same office, hooked 2 keyboards and 2 mice into the same PC and spend literally the entire day alternating who's writing code in the same IDE.

I think the concept was supposed to be about helping developers share their skills. Every developer has an experience of looking over somebody's shoulder and having an "oh shit, I didn't realize there is a shortcut/tool/command for that" moment. And the idea with multi-player programming is to get you to see that and learn from it more. But. It. Doesn't. Work.

The problem with the later is AI’s entire sales pitch is to flatten those definition. If you can type it, the AI should be able to figure it out for the most part. You don’t even have to be very clear in your typing. You can just address the AI casually as you would address a person.

Entertainment apps are about increasing engagement. Like an AI driven game, ARGs, or this site. For example, Sora app was an AI entertainment app but it obviously was the wrong approach. I think there is an untapped potential there, but it remains under-explored because most are focusing on productivity now.

As a result, the problem you solve have to be sufficiently complex with a huge surface to guarantee platform operators don’t cannibalize you. If LLMs weren’t an active area of expensive research and advancement, for example if it was just a couple of models with a simple defined interface, majority of people would use whatever one came on their device or OS. Dropbox had a novelty for its file syncing simplicity and it took few years for platforms to catch up. it was able to establish a market share, but the scale of the others dwarfs it. If LLM advancement hits a wall an all models converge, we might find Apple, Microsoft and Google becoming the most dominant consumer AI providers with other experiences building on top of them, like how apps do today with these storage services.

Large behemoths struggle in innovating new and novel ideas, but they can execute on well defined ones. For example, imagine if something like Tailscale’s use case becomes very common place. You would have “Apple Network”, “Google Chrome Connect” and “Windows Copilot Device Active Link”

That’s what a “tragedy of the commons” is though. The degradation of a shared resource (state capacity) by people acting in their own best interest.

The state using AI “to make bureaucracy instantaneous” is a tragedy of the commons. The experience is degraded because before you trusted your grievance will be heard, but now you might get auto rejected by an AI. It’s pretty much the definition of the phrase.

On Windows, it’s not perfect because it’s ultimately cooperative, but at least the developer of the application can also configure how it should be uninstalled. If you mostly stick to “reputable” applications, you know they would have gotten so many complaints about random things they leave behind that they slowly fix their uninstaller. Also you usually get the option to completely uninstall or keep data files, etc. Though it’s not universal. Uninstaller apps like CCleaner used to be more popular, but I hardly hear about these things anymore and usually advise against them.

On Linux, the story is slightly better as you rely on your distro package maintainer to configure the right uninstall logic. Removing random config files in /etc/ or data files in /var/ or systemd units, etc. Usually maintainers get it right as long as you stick to those feeds, and when something is missed it’ll get fixed eventually.

On Mac it seems to be either google it to get a list of locations to clean (which is what I do), or just delete the folder and ignore the random files or use CleanMyMac ($19.99) or CCleaner ($12.99) or MacAppBeGone (that’ll be $8.99) etc. luckily you can try to stick to homebrew and the App Store, but it’s limiting.

I think that’s a future plenty of people see or want. But simpler and composable Linux commands are far more flexible than pre-defined Windows or Mac “Settings App”. The barrier to entry for the former was knowledge and familiarity, while the latter optimizes for simpler user experience. Agents don’t have to worry about knowledge or familiarity the way users do.

Ok, now you have customers_aggregate(‘topspenders’), but the prompt to the agent was “Get me the top spenders that are public companies and show me ones that increase their spend proportionally to their revenue increase and ones that don’t”. If the agent had access to the needed APIs (or CLIs) for your data and a finance API provider, it can generate a script that crunches down the top 10 or 10,000 for all that data. You could obviously come up with a toolcall that encapsulates that or rely on multiple toolcalls and on the LLM to fill in data correctly between them then do the math using inference.

Edit: even if you look at that customers_aggregate(‘topspenders’) tool, you wouldn’t have needed that if the agent could execute a SQL query. It can also express anything the data can do. You obviously trade the control and correctness guarantees you can give for a `customers_aggregate(‘topspenders’)` vs a potentially wrong SQL query generated by the LLM with the infinite flexibility of a SQL query to be able to express far more than you could in a customers_aggregate call.

All those teams, including ChatGPT and Claude.ai, have figured out that you will eventually need to give your agent a small sandbox Linux environment to unlock the same level of “intelligence“ those coding harness exhibit. Stitching together the results of a cli command through scripting or coding gives the agent a ton more flexibility in what it can do as it can utilize its text generation capability into executable logic. toolcalls mostly work for actions rather than complex and novel problem solving. You are making the agent represent a programming control flow through toolcalls while carrying the context between them in a lossy, nondeterministic, wasteful, slow and rigid way.

It’s one thing if you want to artificially limit that agent to a very strict set of available APIs that it must use in a specific way while transferring context between them through the LLM and you don’t want to incur the cost of the extra sandbox compute. But coding harnesses have demonstrated that letting the agent write a small shell or python script can let the agents solve problems that you haven’t even really anticipated in your toolcall approach or that tool calls make prohibitively expensive or not even possible.

But also the token cost tends to dwarf the sandbox compute cost, so why not pay the $0.05/hour to have a sandbox where the agent can run free when you are already paying orders of magnitude more for the tokens

This was always a vaguely desired thing, but the implementation details was always a form of OIDC federation which is mind-numbingly complex and not worth the hassle. Just shove an api key or a shared secret of some sort and move on. Only accounts that are very high value targets for attacks (cloud infra accounts, CIs infra, billing, and things they interact with implement OIDC federation. But that awesome Thai recipe blog has no chance of implementing something like that.

With Agents driving internet traffic, there is a desire to have identities for them. And since it’s inherently easy for them to leak these identities (look at the million sandbox solutions that all share the same dumb goal of “protect the API Key or password”), having short lived tokens for everything (AWS IAM or Azure Managed Identity like) simplifies the entire process. Agreeing on the single IdP was the problem, there are like 40 (Google, Apple, Microsoft, Facebook, Twitter, even GitHub, and Amazon have SSO not to mention all the regional specific ones)

There is a legitimate need for that, and it looks like cloudflare figured if they are that “internet agent identity provider”, then there is a lot of power and control over the internet and AI use in general.

The only thing I see that’s unique for “agents” when it comes to cloud services is finally paying attention to provisioning speed and developer iterations.

Last year, for the first time in my career, I heard c-suite exec concerned about projects with multi-hour long CI or deployments that take hours because “agents can’t be successful” as they need to iterate way more.